Which Changes Matter? Towards Trustworthy Legal AI via Relevance-Sensitive Evaluation and Solver-Grounded Reasoning

In three linesarXiv study on legal LLM evaluation. Existing models are sensitive to legally irrelevant variations. LexGuard, an adversarial multi-agent framework, formalizes statutes into executable constraints and uses SMT solvers to verify legal satisfaction and logical consistency.

Summary generated by Claude — human-verified